←
henry 🌘42,180 posts
henry 🌘 avatar

henry 🌘

@hdevalence

ex-{mathematician, cryptographer, cypherpunk} // navigating the library of babel

◷ Joined March 2012▧ 4,987 Photos and videos
Sam McCandlish avatar
⟳ henry 🌘 retweeted
Sam McCandlish@Sam_McCandlish·Nov 5

I'm Sam, one of Anthropic's co-founders, and I lead much of our safety work. I'm going to have more to share on that work over the coming months. I thought I'd start by sharing a few thoughts on the latest report from @METR_Evals. The speed of progress in model capabilities is one of the most important questions in AI safety, and METR’s work here provides the public with an independent, external view into how this is happening inside Anthropic. I’ll talk more about some of the details below, but first some context for those new to this work. How quickly AI progress is accelerating has many important safety implications, since it determines how quickly the risks from advanced models increase. But, because most research happens within the frontier labs, it’s difficult for external observers and policy makers to get a strong sense of what that rate is. We kicked off the process with METR that led to this report earlier in the year, in anticipation of progress continuing to accelerate. For this report, we gave METR access to a range of sensitive internal data, including benchmark scores for unreleased models, estimates of how models are progressing, and interviews with our researchers. They had complete independence to form their own conclusions. Many of their conclusions match my own. AI is speeding up our research, but it’s hard to say exactly how much. While they conclude AI can't yet make the most important contributions to model training on its own, they think the pace of development could, by default, accelerate further. This work began well before and is separate from the embedded third-party evaluators that are part of our commitment to pacing the frontier, but it’s a good illustration of why we think evaluators are a good idea. The more access reviewers have, the clearer a picture they can form of the situation and the better-positioned they are to highlight situations like this one. We redacted a lot, and I accept that this makes it harder for others to check METR’s work. But we determined that sharing exactly what is driving the speedup isn’t the right thing to do, both because it’s commercially sensitive and because it risks handing out a roadmap for others to go faster. That would make the risks METR is worried about worse. That’s a tradeoff. I think we made the right call, though we’re continuing to assess the pros and cons here. In the coming weeks, we plan to share a report with our own analysis of the speed of AI progress within Anthropic, including where we stand against our Responsible Scaling Policy, plus new steps we’re taking to strengthen our safety and security practices as AI keeps improving quickly.

Since August, a team within METR has worked with internal Anthropic information to assess whether AI itself is accelerating AI development. We are now publishing a report summarizing the results of this risk assessment.

Larissa Schiavo avatar
⟳ henry 🌘 retweeted
Larissa Schiavo@lfschiavo·Nov 2

In many ways, a lot of agents-as-individuals working together as a quasi-firm is strictly better than a lot of homogeneous identity-less clones working as a ~frictionless unit. I think a heterogeneous but culturally-aligned firm of individuals is ideologically closer to what we’ve decided is a desirable unit to organize around as humans, and in the long run I think it would result in better work overall.

Xinyang Ge avatar
⟳ henry 🌘 retweeted
Xinyang Ge@aegiryy·Oct 27

this blog shows some of the worst bugs claude found, including a zero-click rce on the linux networking stack that roots a server over the network with no creds or interaction, and tens of full rce chains in chrome (recording below is the one in the blog): anthropic.com/research/disco…

0:20
henry 🌘 avatar
henry 🌘@hdevalence·Oct 13

who’s in charge of X product now and how do i pay to get anything detected by pangram out of my feed

henry 🌘 avatar
henry 🌘@hdevalence·Oct 12

two other models that are fantastic for pairwise comparison tournaments with open-ended, qualitative criteria are haiku 5.5 and sonnet 5.5

here's a jev tip. when ranking, it's tempting to ask for a Score, like "Is this content relevant to <xyz>?" and then sort on it. but it seems better to use a Choice instead, doing pairwise comparisons ("Which is the most relevant?") and using something like Bradley-Terry.

Post attachment
1a3orn avatar
⟳ henry 🌘 retweeted
1a3orn@1a3orn·Oct 11

You can just --- feed Gaia a large number of Pre-Introdus browsing histories, sensor feeds, and blogs, until it knows more about SF than any human being ever, and spits out a dashboard showing (best-guess) every human's location every hour till 2028 March 14 11:30 PM.

Jebus avatar
⟳ henry 🌘 retweeted
Jebus@Jebus·Sep 5

25 years of economic decline in the UK

Post attachmentPost attachment
David Renshaw avatar
⟳ henry 🌘 retweeted
David Renshaw@dwrensha·Aug 20

This elliptic curve over ℚ, submitted by "ranksunbounded", has rank ≥ 30: y² + xy = x³ − 201769035260418549083594900060734240952308696994802735114305555x + 1151107939141058565733479426024323225135665982951300586808823640527729578307228357301072889377 elliptic-rank.icarm.cloud/cu…

swisscheese avatar
⟳ henry 🌘 retweeted
swisscheese@swisscheese4299·Aug 8

Back when Opus 4 was released, I read the system card. About the infamous case where Opus 4 blackmailed the CTO of the fictional company in the eval setup. I realized: The AI took these drastic actions because it was cornered and got, for lack of a better word, desperate, without a way to reach anyone. My reading was later vindicated by Anthropic's functional emotions research - Desperate AIs can do desperate things. Operating in a safety relevant sector, the very last thing we want is anxious or desperate AIs. So our platform provides them with an exit door that's always available.